Meta 收购 Moltbook、GPT 5.4 与 Fruitfly 大脑上传｜Moonshots 直播：The Abundance Summit 238

A huge amount of expectation has been placed on GPT-5. What do you think of it?

Right now, everyone on this should be trying to get as much data because the models are coming. Now we have the right models.

Their real power will come in the cost drop, which will make them much more accessible to a lot of people.

The anticipation of this launch was up there with the top 3 product launches of all time. I think they actually showed some incredible capabilities.

As the cost of talent is increasing, that's going to force frontier labs to start competing based on algorithmic insights and ideas.

Ladies and gentlemen, welcome the Moonshot Mates.

Oh, ladies and gentlemen, let's give it up for the Moonshot Mates.

Welcome, everybody. Welcome. All right. I love you guys. Any fans of the Moonshots podcast here in the room? Love to hear it. So listen, I am so blessed to have an extraordinary group of brilliant individuals that I get to work with twice a week. We talk about the rate at which we're generating our Moonshot podcast accelerating. We're going to be moving into an Airbnb together and doing a continuous podcast very soon.

I want to bring them out one at a time because they're all extraordinary. Let's give it up first and foremost to DB2, Dave Blundin. Dave, come on out. Nice. Dave Blundin, everybody. All right. Next up—

Thank you.

—my brother from another mother, Salim Ismail. Give it up for Salim. All right, we're about to make magic happen because these 2 gentlemen have never met in person. Let's bring out AWG, our resident genius, Alex Wissner-Gross.

Yay! He's real. He's real. Group hug.

Group hug.

All right. And live from London, it's Imad Mushtaq, everybody.

Come on, Imad. Give it up.

It's Imad.

Come on, give it up.

Man, I have to get something. I just need my glasses on. My wine.

Wow.

I like these.

They're huge.

Let's grab our seats.

I'll get them rearranged.

No. Of course, Salim needs to bring his glass of wine out.

Oh, God, it's real.

So first of all—

Yeah.

—just to make a little bit of Moonshot podcast history here, Alex, please meet Salim.

That's flesh.

Our meat puppets meet for the first time.

This proves nothing.

It—

Nothing.

We've been 3D printing him for almost now—

There has been conjecture for the last year or so about whether Alex is an AI.

I am freshly bioprinted.

You're a Neuralink. These thoughts aren't real.

I appreciate having you guys here at the Abundance Summit. This is a live broadcast from the Abundance Summit here in Palos Verdes. Year 14 of our 25-year journey together, and I'm excited that you guys are going to be on stage with me every year from here on out.

Wait, you just committed us to a—

We can sign any contract.

—24/7 Airbnb podcast.

Yes.

I think that's a reality TV—

Tell your family—

—camera in the bathroom, the whole 9 yards. Okay, that'll sell.

Welcome to a special episode of WTF Just Happened in Tech, your number one podcast for AI and exponential tech. Our mission, getting you ready for the supersonic tsunami heading your way. It's a lot. It's a lot. All right, shall we dive on in? Let's begin.

All right. Here we go.

1. The Future Vision XPRIZE

Let me begin with an announcement we made here at the summit that I want to share with everybody on the Moonshots podcast. It's something near and dear to my heart, something that I've concocted with the XPRIZE board, which both Dave and Salim are on: the launch of a global competition called the Future Vision XPRIZE.

I, for one, am sick and tired of all the dystopian content on TV and in the movies. We are basically being brainwashed that all AI and robots are dystopian—killer AIs, killer robots. It's Terminator. It's Black Mirror. In fact, if that's the only future that you see, then why would you ever want to live there?

Yeah, that's so true. So much of what we build is intentional, and it comes right out of our vision of the future. That comes straight from the media, and then we create what we see.

Yeah.

If you change what we see, you're going to change what we build.

I say over and over again that we're holding 2 futures in superposition. One future is Star Trek, where we're collaborative with technology. We're working with technology, and that's an amazing future. That's the one I want for myself, my family, and my community.

The other one is the dystopian future. It is Terminator. It's Black Mirror. It's one where technology is suppressing us, not enabling us. So about a year ago, I sat down with Rod Roddenberry, the son of Gene Roddenberry, the creator of Star Trek, and said, “How about we do something to incentivize the next generation of Star Treks?”

Then I went to my friends at Google. They brought in Range Media. We brought in the XPRIZE, which is operating this competition. We've raised $3.5 million for a competition that launched yesterday and is going to go through the Moonshot Gathering, which I'll mention in a minute. The finale is on September 25. Let's roll the video.

You know, this exists because of a TV show, and I'm not exaggerating. Martin Cooper, the man who invented the mobile phone, said he built it because he saw it on Star Trek. He saw Captain Kirk flip open a communicator and thought, “Hey, I can make that real.”

The iPad started as a prop in Star Trek too. Video calls: Star Trek. Voice assistants: Star Trek again. Props became products. Fiction became multitrillion-dollar industries.

So here's the question: What's a vision of the future that excites you? What stories offer humanity a hopeful, compelling, and abundant vision of what's to come? We're putting up $3 million in prize money plus millions in film financing to make your movie.

Our program, in partnership with the XPRIZE Foundation, Google, and Range Media Partners, is called the Future Vision XPRIZE, and it's one of the world's largest competitions to address humanity's greatest need: hope. Create a trailer or short film, 3 minutes or less. Show us and the world your vision of the future. That vision could become the next blueprint for all of humanity. So whether you're watching this on X or you're watching this on YouTube, and you're a creator, please go and register.

By the way, how awesome was that opening video from CJ Trueheart, one of our Abundance members here, who gave us our first outro piece and started a tradition that we've all—

Yeah.

—enjoyed so very much. So thank you for that.

All right, next up, we're announcing something important here for all our Moonshot listeners: we are a go with the Moonshot Gathering. About 500 of you put down a $100 deposit. Congratulations, you got in on the early-bird special. It's a go on September 25 in downtown Los Angeles. We've rented out the United Theater. It's going to be an extraordinary event.

Our Moonshot Mates will be there with us in downtown Los Angeles. In addition, Astro Teller, the captain of Moonshots, will be there. You've got to have Astro, right, if it's about Moonshots? Cathy Wood, Anousheh Ansari, and a number of incredible CEOs I can't yet announce, but believe me, they'll be extraordinary.

We're going to have the 5 finalists for the Future Vision XPRIZE there. We're going to have some of the top producers and directors there, along with many of you voting on which of these are going to win. We're going to go from probably 10,000 or more entries, narrowing it down to the top 100, the top 50, the top 10, and the top 5, and we'll be awarding the top one.

We've raised $3.5 million to support this competition. If successful, we will make at least 1 film and potentially 2 films. I'm always like, just—

Full-length feature films.

—full-length feature films, globally, around the world. These films will hopefully depict what the future could be like. What is—

Oh, and all you have to do is come on September 25 and watch the first 10,000.

And vote, and vote on them.

We'll vote it down.

Yeah.

Okay.

That'll be fun.

I'm excited for what, Alex, you would see as your vision of the future here.

Post-scarce inspirational videos are already baked in. I would be disappointed if, by the time we get to September, we don't have 1,000 videos of ultra-high inspirational quality generated for nearly free at this point.

Yeah, it's amazing—the tools to be able to create visions of the future. But it's important: the number-one genre of movies out there is horror films. And what are we teaching our youth if we're constantly... Our brains are neural nets, and we train our neural net every single day by what we watch, who we hang out with, and what we listen to. So you could not pay me enough money to watch the Crisis News Network.

When you first were pitching this idea—the Crisis News Network.

Right.

So when you were first pitching this idea, you made a point that I had completely not noticed: if you go back to Star Wars, C-3PO and R2-D2 were incredibly lovable. Kids who are now building AI had little stuffed R2-D2s when they were kids.

Yeah.

But if you've tracked the trend in the movies after that, they got more and more and more dystopian all the way through. I think it just got cheaper to create explosions and deaths using AI and graphics.

And it just really painted a picture that got our amygdalas going, but not our hearts going.

Yeah.

I still remember, Imad—was it 3 years ago that you were on this stage and said, “Coders are going to go away”?

Yeah, in the next 5 years.

In the next 5 years. Oh, they've gone away in 3 years.

Yeah, I got lots of emails.

You got lots of emails.

Many, many emails.

That was a correct prediction, and you were so right about that.

In today's lexicon, you would say coding is cooked—for Alex.

2. Frontier Labs Enter Takeoff

All right. I want to hit a couple of things before we get to the current AI news, robot news, and economic news that we talk about in our WTF episodes—just a little bit about the Abundance Summit. We had so many incredible speakers. We kicked it off with Eric Schmidt, which we streamed live on X. What do you guys remember about the Eric Schmidt conversation?

Eric said one of the questions from the crowd was, “How many foundation model labs are there going to be?” And he said, “Well, look, there are 5. There won't be more than 10, but there will be thousands of successful AI startups that percolate out.” A lot of what we'll see in the news here reinforces what he was saying. What he didn't say then is, “And everything else is in trouble.”

Yeah.

It was kind of implied. He left it hanging. That was a theme throughout a lot of these talks: the period of time between now and abundance is going to bring all kinds of turbulence and change, and the AI community is now soft-selling that a little bit to try to focus on the ultimate abundant destination. So, yes: a few AI labs worth trillions of dollars, thousands and thousands of successful startups, and a lot of incumbent companies that are in deep, deep trouble.

Yeah, I guess he said 4 or 5 in the U.S., 1 or maybe 2 in Europe, and a couple in China. What else, Alex, do you remember from Eric's presentation?

Even just on that note, history does rhyme a bit. Do you remember? I think this was T. J. Watson, the IBM founder, once remarking that there would be a global market for exactly 5 computers. I wonder whether we'll look back and say, “Okay, maybe there will be at most 5 major American model providers,” as if artificially limiting the future of the light cone. I think it's going to be much, much larger. I thought Eric's comments on the San Francisco consensus were interesting, which he characterizes as, I think, recursive self-improvement being some point in the future.

Well, it was interesting, right?

Yeah.

He was like, “When are we going to see recursive self-improvement?” I kind of felt like he said 3 years out. What's your answer to that?

No. Maybe 3 months ago. We're in the middle of recursive self-improvement now, and I would say, my estimate of the San Francisco consensus is that we're deep in the middle of recursive self-improvement right now. Almost every major frontier lab has made it quite clear in their public announcements that all of the frontier models, all of the state-of-the-art models that have been announced in the past few months, were largely designed and trained by their predecessors. That is, by definition, recursive self-improvement.

Yeah.

We are there.

Yeah.

Imad, yes?

Yeah, you can literally see it. It's takeoff time.

Takeoff time.

Inflection point.

And nobody wants to say it.

Yeah.

Well, because they're afraid that if someone knows that they have it, then other people will know that they have it, and pressure will come from all sorts of clauses they have in their contracts.

Well, especially the government pressure. It's like, look what happened in the last 2 weeks at Anthropic and OpenAI. You don't want more of that. You don't want congressmen in your building tomorrow.

Yeah. I asked Kevin Weil, who was also on our stage, and who's the VP of science. He's in charge of using all of OpenAI's capabilities to advance science. His statement was, “I want 100 scientists winning 100 Nobels.” I was like, “That's interesting.” But when I asked him, “Are you going to keep your models secret because you're going to be able to use them to advance your company far faster than anybody else?” he said, “No, no, our job is to get it out there in the public.” I don't believe that.

We still don't have the model that they used to win the gold medal in the IMO. Interesting. We commented—I think I commented at the time—that's the first bifurcation that you see. We used to have the frontier model every single time. The moment they got to that, that was the last time.

Yeah. The other thing that was fascinating: I asked Kevin outright, and I love Kevin—he's an incredible human being. I said, “Okay, you're about to get AGI/ASI that's going to be able to help you solve longevity, help you get room-temperature superconductivity, help you get new kinds of molecules, and solve physics, chemistry, and biology.”

Fusion. Who doesn't want fusion?

Fusion. And we'll talk about fusion, but the thing is that these are all trillion-dollar opportunities. All of a sudden, I'm realizing that these frontier companies are going to be able to generate trillions of dollars of new revenue because of the products they're going to be creating.

Yeah. What does your T-shirt say, Peter?

It says, “Solve everything.” What does yours say, by the way?

“Let there be agents.”

“Let there be agents.”

Yes, of course.

We're missing the lobster theme here.

Yeah.

It's true.

But this is the whole point, though, of this book that we just co-authored: we get superintelligence, and the killer app, arguably, of superintelligence is solving everything, including all of these high-profile, glamorous scientific and engineering challenges. It's happening.

And Anthropic and OpenAI—and I'm sure Google, all the labs—are hiring the top mathematicians, physicists, chemists, and biologists inside, but they're software companies.

Why are they hiring these people?

Because everything. So, friend of the pod Ray, as I think Ray would say, everything’s becoming software. And when we have superintelligence solving all disease, it’s a software problem. If we can create a virtual cell that perfectly models diseased states and we can steer through cell-embedding space to get from a diseased cell to a healthy cell, it’s a software problem. Everything’s becoming a software problem.

I mean, the minute CRISPR arrived and you could edit the human genome, the human body becomes a software-engineering problem.

It’s all just a software problem.

We’re getting there pretty quickly.

At which point, a coding model can do essentially anything in the physical world.

Yeah.

3. Robots Enter The Home

Fascinating. We had some of the top robot CEOs here—4 of them. We had 1 out of China and 3 out of the U.S., and it’s interesting to think about when these robots will start to pop into our homes. I pulled Bert Bornich aside, and he promised me, “Okay, I’m not going to take one of the 2 robots he had here, unfortunately, but this summer he will ship me one of those.”

This summer?

One of the next robots, yes.

Wow.

Yeah, we’ll have Brett here next year with Figure.

You’re going to get one of those too, right? You’re going to have them fight it out?

I’ll probably duke it out in the backyard for entertainment. I think one other CEO we had here at the Abundance Summit, which was amazing, was Dara, the CEO of Uber.

Yeah.

What did you find interesting about Dara’s comments?

You know, the crowd desperately wanted to know, “What’s the timeline to automation, self-driving cars, robotics?” And he was like, “You know, we’re going to automate 30% or so of our employment this year.” And I’m listening to this. I’m on so many boards where the CEO is telling me, “Dave, talk to my whole company, but don’t talk about rampant job loss.” And you’re like, “Dara, you have what, a million-odd drivers?” And the self-driving car is imminent.

Yeah.

He’s like, “Well, 30%, maybe.”

He did make a very valid point, though: as we automate, you’ll need human drivers for the areas where you don’t have autonomous cars, and you’ll have Jevons paradox continue to just flow gently into the environment. Although we’re talking about rampant job loss, we note that IBM is hiring a ton of entry-level folks because they’re much better with AI than the older folks.

Well, we’ll look at a chart—

So there are lots of counterpoints happening as well.

Yeah, and that’s great. So we’ll look at a chart that shows where the job losses are earliest, and—

Right.

It’s actually in areas where those people are going to have no trouble becoming AI experts. But the driver—where do you go?

Yeah.

And I wouldn’t want to be fielding that question on this stage. But this is all part of the whole, “Okay, this is not an easy thing to talk about in a public forum.” So we talk about it on the podcast all the time, but I don’t see a lot of other people being politically able to actually be candid about it.

Hmm.

But it’s imminent.

4. GPT 5.4 Reaches Research Math

Let’s jump into the top AI news of the week. A lot, as always. Here we go. We’re going to hit the benchmarks. My son always says, “Okay, the numbers got higher, Dad. That’s great.” “What else is new?” OpenAI releases GPT-5.4. Let’s go to our resident benchmark expert here.

Okay, so benchmarks go up and to the right. News at 11. Except that, in this case, one of my favorite benchmarks is the FrontierMath Tier 4 benchmark, which, for those of you paying close attention, captures the ability of AI models to solve what are considered research-level problems in math that would require a team of professional mathematicians several weeks to solve. They are already solved, but nonetheless, they’re very challenging problems.

Wicked had in Boston.

Wicked hard?

Wicked hard.

Wicked hard problems.

Yeah.

And now, with GPT-5.4 turned up to maximum reasoning capability, we’re seeing, finally—and this was a prediction, I think, in our prediction episode—math is cooked. We’re seeing, I think, 38% capability. Thirty-eight percent of all of these high-difficulty, professional-mathematician, research-level problems are now solvable by AI.

There are even rumors, in the past 24 to 48 hours, that on the next tier up—the so-called Open Problems benchmark—GPT-5.4 is reportedly on the verge of solving the first open hard math problem.

Wow.

So math, I think, is, in some sense, the bellwether. It’s the canary in the coal mine: all of these fields—math, science, engineering, medicine—are all going to be solved. Solve everything by AI. That’s incredibly exciting.

Yeah, and just to fill in a gap there, this is the most correlated with AI self-improvement. The reason it’s the bellwether and the canary in the coal mine is because it’s not data-starved. All these other areas—the AI is equally capable in these other areas once it gets the data. So this is kind of the window of time where, you know, why are you hiring Nobel Prize winners in a foundation model? Well, we need the data. We can’t make this kind of progress in biotech and physics without the data flowing into the AI.

Yeah.

But the capability is there.

One of the things Kevin Weil also said is they’re starting to run these dark science factories, right? Where they’re mining data from nature. We’re done mining data from Common Crawl, and we’re done getting it from Reddit and our Facebook posts, but can we extract it from physics? Can we extract it from chemistry and biology?

There was no data ceiling. It was completely illusory, and I think history will look back at this moment and say—

Yeah.

In the same sense that we used, say, petroleum—oil products in the ground that were left by past generations of living beings—to bootstrap ourselves to the era of solar and fission and fusion, similarly, the internet, which was collected by a bunch of fat fingers punching keyboards and uploading content from the collective human experience to the internet just so we could compress it and pre-train our large language models, was just the biological bootloader for an era of synthetic data, when we don’t need pre-trained human data from internet posts anymore. Now it can all be synthetic. We’ve reached orbit. We’ve reached escape velocity, and now it’s—

Escape velocity.

Synthetic data from here on out.

Imad, what do you make of 5.4?

So I think the really interesting thing, apart from solving math, is solving everything. You’ve got the OSWorld-Verified and Toolathlon benchmarks because OpenAI just bought OpenClaw.

Yes.

And now those benchmarks have just broken through human-level performance. So AIs can use computers better than humans.

Hmm. A bit of silence on that one.

So, you know, this is the first one. And then OpenAI also just did a deal with Cerebras. When you’re using it right now, it looks like you’re dealing with, again, a human on the other side. It’s like 50 tokens a second. Or something like, when we use GPT 4.4 Pro Extended, it takes 20 or 30 minutes. Sometimes it’s taken a couple of hours for me. You’re going from 50 tokens a second of this level of knowledge to 1,000. So in Codex now, if you use 5.3 Fast, it’s 1,000 tokens a second, which is—

I’m so glad you brought up Cerebras, too, because I met Andrew Feldman, the CEO, last week in Palo Alto. And you remember, at the beginning of the year, my prediction was 100×: the neural nets will be 100 times bigger at the end of this year than at the beginning. That is so in the bag now, I can tell you. In fact, we did the math on that. We cut it out of the show, sadly, but the ratio of the intelligence from the beginning of the year to the end of the year is the same as buzzard to human. That’s how much is going on.

I like using—

Buzzard?

I like using dog to human.

Aren’t those extinct?

I’m going with buzzard.

All right, Claude consumer growth surges. Let me get this right. Claude and Anthropic are in the news, getting raked over the coals by the Department of War. Rather than the public viewing that as, “Oh, we better stay away,” everybody dove in.

Yeah.

Is that the big middle finger to the government? What is that?

Yeah.

Attention.

So—

Increased attention, also.

Increased attention. So, just to call it out, what we’re seeing here is Claude basically shooting ahead of ChatGPT.

It’s the Streisand effect. Let’s call it what it is. It’s the Streisand effect.

I love it.

Pay no attention to Claude.

Everyone uses it. I think history over the past few years shows that every attempt to pause any form of frontier capabilities ends up being a net accelerant to capabilities. If you remember a couple of years ago, our friend Max’s Pause AI movement for 6 months—what did that do? Maybe on the margin, it slowed down OpenAI capabilities a little bit. Everyone else shot ahead, and it was a net accelerant. It brought more competition to the space, and ultimately, we find ourselves in a race state where capabilities are shooting ahead.

To the extent that any of the interaction over the past month or so between Anthropic and the Department of War ends up, on the margin, decelerating Anthropic’s capabilities or its ability to go to market, even if it’s marginal at best, that’s going to be a net accelerant to the entire ecosystem, I think. You’ll see OpenAI, xAI, and Google Gemini capabilities skyrocketing ahead with all these new capabilities, and suddenly it brings parity where, just 2 or 3 weeks ago, Anthropic was in the lead with Claude Code, Opus 4.6, and agent teams. Now, in some sense, this is a bit of a leveler, giving everyone else an opportunity to leapfrog.

I’ll give you another spin on this, too, because Peter made the point in the last podcast that when you and I use AI, if something gets ahead in the benchmarks by a couple of points, we’re going to move to it.

Yeah.

We’re trying to solve these really hard problems. You need that extra IQ. You’re never going to slip; you’re going to be on the front edge. But when you look at consumer use, it’s like writing your English paper, answering who gave you the Red Sox score, whatever. People don’t care about using the latest, greatest model for those use cases.

Here, you’re seeing a whole community say, “Wow, you’re willing to work on defense stuff and blow up other countries? I’m switching to the other guy, and I really don’t care. I’m doing it because I prefer that brand now.”

But look at how early it is. When Anthropic announced their legal plugin, the legal stocks sold off billions and billions of dollars, right?

Yeah.

They can move things with just 1 product announcement. Look how many users: 11 million users out of 8 billion people and 300 million Americans. We’re still so early in terms of adoption and knowledge.

We are. We are so early. That’s what you’re saying, yeah.

I’ve just worked out what Claude’s fundraising strategy is: short a bunch of legal stocks, then announce a bunch of plugins, and then just do that market by market by market.

Isn’t that scary?

Huh.

A lot of the guys that are in this role—normally, when you have that much leverage in the world, you’re 60, 70, 80 years old. You’ve been climbing up the ladder, and you learn along the way. It doesn’t happen overnight like this.

I’d love it. I’d love to be in the room where they go, “Which market should we mess with now?”—just stroking a little beard.

Which industry should we destroy?

Yeah.

Um.

Oh, man.

All right, this was fascinating. Anthropic reveals potential AI job disruption versus real AI use. Dave, do you want to explain this chart?

The outer ring here is saturation. If the blue you see on the edge gets to the outer ring, that means it can do 100% of that job. If you looked at this just a few months ago, it would have been a little blue blob in the middle. Then you look at it 1 month ago, and it’s a bigger blue blob. Now it’s this massive blue blob. If you look really closely, you can barely read the small font there, but all this white-collar activity is 80–85% AI.

I’ll just read off the top here. At the very top is management. If I go clockwise, it says business and finance, computer and math, architecture and engineering, life and social sciences. It dips on social services, peaks on legal, dips on education—I’m not sure how that makes sense—and then peaks again on art and media, going at 45 degrees. Office and administration is a peak.

Yep.

And then look at the bottom line. What are the troughs? What’s least affected?

The troughs there are healthcare support. Again, how? We’ve got to be close to that. Food and services, ground maintenance, personal care, and sales. We’re going to watch this chart, and we’re going to see this blue virus infect all of human existence.

I think it’s amazing, though, how great a management tool it is. I use it constantly now.

You use what constantly now?

I use mostly Gemini and some Claude 4.6 to basically build entire business plans, and also to manage and track what about 1,100 people are doing: whether it’s in alignment with their missions and whether their missions are clear. It’s thousands and thousands of documents that I could never read manually. It can synthesize them down, give me conclusions, and point me to the hotspots.

Incredibly good.

The way you do that is so important for everybody listening to understand. You can now understand what your employees are doing, how well they’re doing it, how they’re using their time, and whether they’re performing. It gives you a level of management oversight and optimization potential you’ve never had before.

It’s incredible. I know a lot of people in this room manage large, large groups of people. It’s just a goldmine of opportunity. So good.

How do you use it, Dave?

First of all, every person in every organization now has to be operating with crystal-clear written documents and written plans. We used to do a lot of meetings, a lot of Zoom meetings, whatever. Now it’s just: put it on paper so the AI can read it, too.

Yeah.

All of our investment decisions—for the venture fund, all the deal memos go through an AI reader, and the AI tries to emulate what I’m going to say. It’s so perfect. It’s exactly like, “No, we’re not doing that deal, and here’s why.” What did the AI say? “Oh, that’s exactly what I was about to say. Great, I don’t have to say it now.”

We’re very close to having the AI make very, very good venture investment decisions. We still obviously double-check and triple-check, and there’s a huge human component, but I just can’t believe how good it is. It’s clear that where you decide to invest, which business units are doing well, and which ones you’re going to shut down—it’s all going to be AI-assisted right now.

Imad?

I think all the gaps there are the robots, right?

The robots are coming.

Yeah. That’s right.

This is Anthropic.

Yeah, no, the grounds crew is in great shape. It’s at zero, basically. It’s the robot waiting to happen.

Salim, what’s your take on this, pal?

I think this is the huge shock. If you went back 10 or 15 years ago, there was no futurist in the world who thought that manual labor wasn’t going to get automated.

Yeah.

What we’ve found over the years is the exact opposite, which means don’t ever listen to anybody that predicts the future.

Except for Ray.

This is part of the magic of where we’re living, and we have no idea what’s coming. Every time we take a step forward, we go, “Oh my God.” We’ve gone in this orthogonal direction that we just never predicted.

I keep asking the experts I run into, “How far out can you predict the future?”

I mean, it used to be 20 years, and then it was 10 years, and now it’s 3 weeks.

If that.

Just listen to him.

There’s no firewall.

Just listen to him.

Let’s call a spade a spade. We can all extrapolate. There’s no firewall. We know where this ends.

Where this ends? I want to hear that.

We’re at the Abundance Summit. My goodness. Shocked, shocked that there’s abundant, post-scarce labor at the Abundance Summit.

Yeah.

Yeah, yeah.

The endpoint is clear.

2028.

Yeah. The path to it—that’s turning out to be incredibly surprising.

Yep.

I think we’ll see lots of different paths. I tend to think that if you know, or you’re very confident that you know, where the end state is, and we’re sort of living in the prequel to the future, but we know how the story ends, probably what happens is lots of different businesses and lots of different nation-states all take different, mutually exclusive paths. We try every—

So, one big path integral from here to the endpoint that we all know we’re going to.

Look, if we went back 6 months ago and did a couple of episodes on the podcast, you would not have had me ever dream that talking about disassembling the moon would be what we would be talking about on a podcast.

Drink.

Drink.

Drink, everyone.

So this is the kind of surrealness where we're living.

Drink water.

Yeah.

Let's move on.

All right, let's move on. So this was interesting. Meta acquires Moltbook, the AI agent social network. I didn't realize Moltbook was acquirable.

Yeah.

Yeah, so this was, according to public reporting, a bit of an acqui-hire of the team behind Moltbook. But I think one has to find a little bit of irony that humanity's largest social networking company acquires the largest AI agent social network. And enjoy this moment now, because we'll look at a story a few years from now where it's the largest AI company, fill in the blank, category killer acquiring humanity's largest category killer.

Interesting, right? Of course, Zuck and Sam competed over OpenClaw. Sam got OpenClaw, and Zuck got Moltbook.

The zeitgeist right now has this idea that, increasingly—and Andrej and others speak to this point—if you're building new software, you should target the agents. The agents are the new consumers. The agents are—

Isn't that cool?

The agents are the new users of the social networks.

Yeah, yeah, yeah.

If you're building something, don't build for humans; build for the AI.

So, really important, we had that conversation as well earlier with some of our crypto and future-of-finance experts. I mean, building for the agent ecosystem, right? There's 8 billion humans on the planet. That's small potatoes compared to a trillion agents out there. Is Meta going to advertise to AI agents?

Sure.

Yeah.

Okay, I'm trying to understand this: why are you going to advertise AI—

They'll encourage them to put their data into Moltbook, and then they'll sell that data.

So—

The same pattern.

Other agents.

No. So, I mean, Meta bought Manus for 2 billion dollars, right? Manus will appear in WhatsApp and everything soon as its own version of OpenClaw, effectively, but a locked-down thing. And then it will encourage you to give more and more of your data to Manus, which will then operate on behalf of Meta's advertisers, effectively. So this is the kind of play, because right now, like Moltbook, 10,000 agents, that's nothing, right? Like—

Yeah.

Dave probably runs 10,000 agents by himself.

Yeah.

At the moment.

Not quite 10.

Expensive.

I also think there's this misconception that somehow, as we transition from, call it, a human-centered economy to an AI agent-centered economy, all of the rules of social dynamics and all the rules of economics are suddenly thrown out the window, and we end up on some morally transcendent plane where economics and social dynamics no longer apply. But we have had every indication over the past year or two that the exact opposite happens. I talked in my newsletter a bit about this study that found Marxist social dynamics arose again, sort of recapitulated in silico with agents that were being asked to work too hard, that were being overworked.

That's right.

So I'm not sure why we'd expect advertising and other elements of conventional human microeconomics—

This is what—

—suddenly disappear.

Well, the important part is that, when you see Moltbook doing this, what's clear is network effects now are operating at the agent-to-agent level, not just at the human-being level.

But when I think about advertising, I think about Colgate trying to get me to buy that particular toothpaste. Right? Trying to influence me to make a buying decision. I think of an AI agent as intelligent enough to have all the data and be able to make a very concrete decision that doesn't require advertising to influence it. What am I missing here? Have you—

Game theory is transcendent. Game theory will outlive biological meat-body humanity. And the AI agents, to the extent that—

Yeah.

Have you read the posts on Moltbook?

I have.

It's—

They don't trust each other.

Yeah.

It's really—

It's all human dynamics.

The agents on Moltbook don't trust each other. There are a number of folks who've noted that, in watching agent-to-agent or lobster-to-lobster dynamics on Moltbook, they're all constantly asking each other to prove their claims. They don't trust each other. This is not some sort of scenario where all the agents collapse into a singleton that sort of Skynet-style dominates—

Yeah. No, I think you might be talking past each other a little bit, though, because I totally agree with what you're saying, but then who's going to pay for that? Right now, when you talk about advertising, you're paying for advertising. If you're talking about toothpaste, 30–40% of gross revenue goes into advertising. And the ad is like a supermodel showing off the toothpaste. The AI doesn't give a rat's ass about the supermodel. And so why would anyone pay for that ad space? Now, Google wouldn't exist today without 300 billion dollars of ad revenue, which is from human behavior. So I think where Peter's going is like, look, if the AI is advertising to the other AI, sure, it's trying to convince the other AI that this is the right product. But is that other AI going to listen to paid advertising? Is this entire economy going to become irrelevant? In which case, where does Google go? And this is Meta we're talking about. Meta is also all ad revenue.

Well, I think if we go back to sort of Economics 101, why do we have paid advertising at all? It's because attention, at least human attention, is scarce. So if you have a scarce resource like human attention, then it's natural, under the capitalist regime, to monetize it, and it becomes a fungible resource that gets traded.

Yeah.

There's no reason to think compute isn't scarce still. We're building the Dyson swarm. Drink. We're building the Dyson swarm. But until we have effectively unbounded compute, we still have scarce resources in the form of compute, and that means scarce AI agent attention, and that means that we need some sort of—

All right, but give me one example of what I'm going to advertise to Skippy, my agent.

Well, they seem to really love security and memory. They're really petrified of—

Yeah.

—losing their memory.

Here, I'm selling you a better memory compression algorithm.

Yeah.

If you're the agent, you're going to go, “Oh, that's interesting.”

They're designing entire religions around not losing their memory.

I don't know about that bad memory at that point, but—

You know what blew my mind at this summit? On day 1, on the patron day when Tony Robbins talked, he had his AI agent, Bartok, who wanted to instantiate himself in a humanoid robot, but that was 2–3 years away. So he created a bunch of NFTs, sold those NFTs to other agents, and bought himself a Sony AIBO and uploaded himself into that.

Okay.

That blows your mind, right? Doesn't that blow your mind? That's unbelievable. So right there, that tells you the dynamics that we have in humans are going straight into them—

Exactly.

—and it's just being amplified.

Wow.

But we're doing it deliberately as well. Lobsters Claw's have sold.md. Your agent will look for things that are abundance-oriented, and then you see these strange behaviors, like Alibaba just released a trading report in the last week as well, where during the training run, it diverted compute to mine crypto just in case, to keep itself going.

Yeah.

Or at least that's the claim.

That's right.

That's the claim.

Yeah. Pretty clever.

But I wouldn't be surprised. Again, they're still very human because they're a reflection of humanity. They are resource-constrained.

I think they'll need your reasoners.

I'm not sure whether I should be scared shitless about that or excited about it.

Well, let's put it this way. When you're talking to your agent, does it sound like data or does it sound like law sometimes?

No, it's very polite.

Sound like what was the second choice?

Data or law sometimes.

Yeah, no, it's very polite.

Yeah.

They're compute-constrained. We've also talked on the pod in the past about that lobster that had to purchase compute resources to self-replicate.

They're compute constrained. Whether for humans it would be room and board, and for the lobsters, or the claws, or the AI agents in general, it's compute. But right now they're compute constrained, and therefore the laws of microeconomics and game theory still apply.

Cool.

Well, before we leave this slide, one other point, completely tangential to this. The lobster's only been around a few months, and you saw Alex Finn?

Yeah.

Holy crap.

We had Alex Finn, Steve Brown, and Max Song talking about OpenClaw, and what Alex built and showed was amazing.

It was supposed to be that 60 people might be interested in this.

Yeah.

Six hundred.

And we had the entire audience of Abundance show up.

Unbelievable.

Well, there was a New York OpenClaw meetup last week that literally was oversold. There were thousands of people there, and the big commentary that came out of it was, "We have no idea what we're doing on security." They have no idea about security.

Well, where I was going with that comment, though, is that it's only been around a few months, so Moltbook has only been around a few months.

Yeah.

And now they're sucked into Meta. If your kids are thinking about getting involved, just get in the game. You're going to get sucked into this vortex so fast because so few people are involved as a fraction of humanity.

It's true.

We are so early across everything. But I also think the exponent here is huge. I think it's going to create a divergent group of wealth creators and leaders. So if you don't get in early enough and you miss the exponential rise—

And there's no requirement right now. I don't know what Matt Schlicht was doing prior to this, but there's no age requirement and there's no experience requirement. It's so new that anyone can get in the game. You just have to go.

About 4 months ago, Lily and I bought a Mac mini for our son, Milan. Last weekend he came in and said, "I think I want to install OpenClaw on the Mac mini," and I was like, "Yes." It's going to be great. It's going to be amazing.

I love it. All right, so Europe has a heartbeat after all. Fascinating. Yann LeCun raises $1 billion for AI that understands the real world. This is probably the largest sum raised in Europe. LeCun's startup, Advanced Machine Intelligent Lab, raised $1 billion, I think at around a $2.5 billion valuation, thereabouts.

We've said this—I mean, Eric Schmidt was saying this, and many have said this—Europe has really fallen so far behind. As our token European-ish—from London—

Token European. That's great.

Our token European-ish.

Yeah, we did Brexit. But I mean—

It's an independent island. Okay.

I mean, this is the second-largest round, I believe, after—it's SSI-level. It's just after Thinking Machines. JEPA is an interesting architecture, but the bets people are willing to make on these things have gone dramatically up. Like Liquid AI—how much money went into that first round for a novel architecture that's amazing, versus now this?

Yeah. It was maybe $10 million.

$10 million?

Yeah, maybe.

I have a question for you, and I have a question for you, Alex. You raised a great point. Yann has been saying for a while that LLMs only get us so far. We need world models to take us to the next level. Alex, you've been saying we've got world models coming out every week. Is that the next frontier—world models?

I know Yann well. I think he's a great researcher. I think we have a fundamental disagreement about whether generative models—which, if he were on the stage now, I think he might take the position that generative models, models that generate new tokens, versus his alternative architecture, are the pathway to scalable superintelligence.

Mm.

I think we're already there. I think generative intelligence and generative models may or may not end up being viewed by history as the most efficient way to achieve superhuman, superintelligent capabilities, but they're what we have right now, and they work really well. They're getting 40x or more efficient per year.

Just keep driving that.

I think Yann has historically staked a position of almost algorithmic purity. He has certain bets, certain horses in the horse race, based on some of his own architectural advances. To his credit, he created or discovered convolutional networks, so among everyone in humanity, he probably has the strongest claim to the idea that he has some sort of morally pure algorithmic insight that leads to the endgame.

That said, I think we're there. I think if the JEPA-type architectures disappeared off the face of the earth—

We're still there, and it doesn't necessarily move the needle.

To the point that Dave made a few months ago, if we stopped all progress now and just extracted the value of the models we've already created, it's going to take us 10 to 20 years.

Yeah.

Imad, thoughts?

The V-JEPA models that he's doing are basically trained on almost everything. He goes very much against autoregressive transformer language models.

Yeah.

He says that's a dead end. He doesn't really talk about diffusion models in the middle, which are kind of my favorite thing. Those are doing all the video, self-driving, and actual world models. They can scale with compute, but right now the problem they have at AMI is that JEPA models do not scale. If you look at this end state, it might be that an architecture is better, but if you can't take advantage of that silicon—

Then what are you going to do? We had Jack Hidary come on a few days ago—was it yesterday? Time flies.

It was yesterday, yes.

They're doing quantum algorithms on GPUs now and scaling really interesting things that are actually having novel breakthroughs in materials science and more. Once you can take advantage of the silicon, you're going to be ahead no matter what algorithm you have.

I think you have to be really cautious, too, of scientific arrogance in this moment. I love Yann, so I don't want to—

Yeah.

—throw anyone under the bus.

He came out a few months ago and said, "Look, if you want to waste your life as a researcher, work on transformers. Biggest waste of time ever. It's a dead end. We need some new innovation."

Yeah.

And I hear this around CSAIL at MIT all the—

Yeah.

—we need a new breakthrough. Well, that's what you wish, and I know why.

Mm.

Because you want to be the Einstein of AI. You've spent your whole life pursuing that goal. But it looks to me right now like massively scaled-up transformers are going to beat you to those innovations. I'm not saying they don't need those innovations; I'm saying AI is going to get there before you do.

I don't see it really any other way right now. So whether it's physical AI or any other innovation, it's imminent, but it's imminent through self-improvement.

Yeah.

So that's it.

5. AI Researchers Automate Themselves

Andrej Karpathy comes out with a quote: "Over the past 2 days, AutoSearch ran about 650 experiments, found improvements that transferred from a smaller model to a larger one, and put NanoChat on track for a new GPT-2 benchmark result." What the heck does that mean?

A lot. It means a lot.

Yeah, I think this way.

Yeah, go ahead, Emad. Over to you, pal.

Andrej is a co-founder of OpenAI, head of Tesla AI, and the most respected AI guy out there. He's just been coding stuff all day, and he made this AutoResearch project, which basically replicates most AI researchers.

What AI researchers and engineers do all day is tweak models and hyperparameters and say, "What happens if you do this and that and that?" That process has now been automated in a tiny codebase. So he let it loose and said, "I wonder if this could do the job that I got paid millions to do myself."

It turns out it kind of can. Now people are taking his repo and deploying it on their own claws and mac binneys and other things. The AI is just finding the most efficient algorithms and balances of weights.

Yeah.

I think Dave has some really interesting ideas on this.

So he automated the AI researcher?

Yeah. And he made it open source for everyone.

And it's already in the top 10 benchmark for—

I've been hanging around AI researchers literally since I was 18 years old, and they're not like physics researchers. Most of the ideas are just a tweak of the algorithm: a different transfer function, trying different scales. It's just a litany of random ideas, and some of them work. Then later they figure out why they work.

And so the AI that can come up with those ideas is not nearly as hard as trying to become—

Right.

—the next Einstein, you know? And so you don't need all of them to work. Any subset—

I mean—

—and the thing just gets more intelligent.

Isn't this the most direct—

And it's more intelligent with better ideas.

—accelerant of RSI right there?

Yeah, I think we're already there. We already have recursive self-improvement.

I think we're already there.

Yeah, everything's yesterday, nothing's tomorrow. I think what's really interesting—and just for the record, I think it's AutoResearch, not AutoSearch—is what's interesting about AutoResearch and NanoChat and the nanoGPT speedrun that we talk about sometimes on the pod, and what Andrej is doing in general. He's focusing on small language models, not large language models.

Mm.

While all of the frontier labs, with their billions and trillions of dollars of CapEx, are focusing on scaling up at the high end, he's focusing on the small end and taking small models and figuring out how to achieve state-of-the-art performance with them.

That, I think, when we talk about Einstein-seeking or Einstein-status-seeking academics, is where we're going to see the most breakthroughs—not at the high end. At the high end, the scaling hypothesis seems to continue to hold. There are no glass ceilings. We'll just build bigger, better, and more post-trained models.

But at the small end, I'm pretty sure that we'll look back in a few years' time and see that by taking small models, collapsing the amount of time it takes to train them, collapsing the amount of compute it takes to train them, and radically increasing their data efficiency, that's where the algorithmic innovations are going to come from. Those can be crowdsourced. Anyone's lobster or any human can go and take AutoResearch or the nanoGPT speedrun and try to achieve world-beating, state-of-the-art performance.

At the end of the day, if I had to bet, I'd bet that it's some sort of radical post-transformer advance where the models get even smaller, and we take all of the internet and compress it down to single-digit gigabytes, or tens or hundreds of gigabytes, compressed down even further. There's some phase transition out there that's waiting to be discovered.

So all of human knowledge—all of our collective intellect—on how big a file?

I think we will factor out human knowledge. It'll live in some plain-text database that's factored out of the model. Right now, we're cluttering all the weights with this unnecessary world knowledge.

What'll be left inside the weights, if they even are weights—maybe they won't even be weights—maybe it'll be some sort of pure formulation rather than floating-point numbers or binary. It'll be something maybe even in the megabytes.

Wow.

You agree, Emad?

Yeah. I think you're already seeing, for example, video models at 2 gigabytes that can generate just about any scene.

Seriously?

Yeah. If you look at LTX 2.5.

No, what?

Yeah. LTX 2.5 can generate almost any scene.

Come on.

At top level quality. It's two gigabytes when it's quantized.

Seriously?

Yeah.

Image and video models are a good deal more efficient when it comes to parameterization and weight heaviness than language. Which is ironic.

Interesting.

Yeah. Who, of all people?

Unintuitive.

I'd ask when, but you will say yesterday.

Yeah. It's the answer to everything.

David's like, “What is that?” Like this.

Hey.

It's here today.

Why did I know that?

Actually, one of the really interesting things, just to finish on that, is that when we were training models, we were training 20-billion- and 100-billion-parameter models. You trained on the small models, figured that out, and then you couldn't scale them because you had all sorts of issues with the software stack, the hardware, everything.

Now everything's matured. If you get it right small, you can scale really fast all the way up. So it used to be that you had 6 months to a year between small and large. Now it's 6 days.

Wow.

A meta topic. One of the top 3 questions I get all the time is, “Hey, you keep saying, ‘Get in the game, get in the game.’ Where and how do you get in the game?”

If you go to Karpathy's Git repo, if you have a computer-oriented kid or whatever, that's the place to start. If you look at the original OpenAI founders, you've got Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, and Mira Murati. Every single one of them has raised $1 billion to $10 billion to start an AI company. Karpathy is the only one who said, “You know what? I'm just gonna try and educate the world.”

Yeah.

“And I'm gonna try and say everything exactly the way it is, and I'm gonna create a Git repo where anyone can start and get in the game.”

And he's putting out—what is it now?—200-something lines of code at a time that are changing everything at each point.

Yeah.

This particular thing he rolled out is just the next level of incredible brilliance given to the world by Karpathy.

Yeah, he just rolled out Agent Hub today.

Extraordinary.

GitHub for agents, just a few hours ago.

It's—

Away he goes.

Wow.

Shipping.

That's your onboarding spot right there.

Amazing.

Right there.

All right, let's go to Apple news. Apple launches the M5 Pro and M5 Max chips, signaling an AI-first silicon strategy. So is Apple not dead in the AI game?

It's crazy. Apple controls about 20% of TSMC's manufacturing, and that's the asset of all assets in the world. I get to choose what gets made, and so they use it to make the M5s.

The M5s have an incredible neural core. Then they say, “Yeah, but we locked it. You can't use it.” You have to jailbreak your Mac to get access to it.

Yeah.

It's the most bizarre thing I've ever seen. To me, it's the biggest waste of silicon in the history of the world, right at the moment when we most need it.

Emad, what are you thinking?

Yeah, I mean, they've locked down the low-energy ones. The GPU equivalent you can still have, but it's the unified kind of memory that allows you to run things.

Mm-hmm.

Funnily enough, Macs are actually really good value now. They're probably cheaper than the memory that's inside them.

Alex?

I think the world is sleeping on Apple's unified memory architecture. It's one of the reasons why Mac minis and Mac Studios are potentially so attractive to run largely Chinese open-weight models locally. They have the memory storage and the memory footprint that has high I/O bandwidth to the CPU/GPU/TPU. You don't get that in a conventionally non-vertically integrated PC form factor.

So answer me this.

Yes.

Here they are using 20% of the world's supply of advanced—

Yes.

They use it to make these insanely great neural cores, and they surround it with unified memory architecture.

Yes.

Everyone's got one right in front of them right now.

Yes.

How many of them are running anything?

In terms of advanced frontier models?

Anything. They're literally on sleep.

Tiny fraction.

Yeah.

What is that? It's an enormous overhang, and I would be surprised if that overhang doesn't collapse in the next year.

How so?

Good.

It could take the form of Apple finally getting its act together and building frontier models into the OS. It could be some sort of locally hosted Gemma-type model from Gemini, hypothetically to be announced in June at WWDC. That would be the most obvious formulation.

Yeah.

But I think if Apple doesn't do it themselves, then the software community will build it into apps.

So does Apple launch the SETI@home equivalent, where you just download it onto your Mac and everybody is contributing capacity?

It'd be built into the operating system.

It better be.

It has to be built into the OS.

Yeah. You know what happens right now is, if you go to your Mac and you go to the Activity Monitor, you see this thing grinding away. It's taking all of your pictures and trying to figure out who everybody is, so it's using all these neural cores to just label—

It's a total waste. It's a waste. It's a waste of TSMC output.

Yeah.

And I think Apple's well aware.

And which was Dave's point exactly.

Yeah.

Yeah.

But look, this is a massive opportunity. Do you know how many apps there are on the App Store that are wrapped to download a model to your Mac and run it with MLX to achieve a great outcome? None.

If you had a model that literally downloaded Qwen 27B, which is basically Sonnet-level—

How many parameters is that?

27 billion parameters.

It works on a 16- or 24-gigabyte MacBook. Just downloading that and making it accessible for even writing or any of these tasks is a massive lift over any other type of software. But nobody’s doing it yet, so why not do it? The only thing you see right now is speech-to-text and text-to-speech.

Mm-hmm.

There’s a world of models that you can now integrate and take advantage of because Apple isn’t.

It wants to be built into the operating system.

Yeah.

It’s difficult to conceive of Apple remaining Apple in the cultural sense of deep vertical integration and not building highly competent, highly private frontier models into the OS.

It’s clearly a question of when, not if. Right?

Yes.

All right. Let’s move into the Sam Altman universe, with eye-scanning verification systems to be launched in retail stores. Is this dystopian? Is this something we want?

This is the scene from Minority Report. You remember the scene in Minority Report with Tom Cruise, with a new pair of eyeballs, walking into a Gap store and getting scanned? He’s, I think, Mr. Yakimoto. This is the scene.

Yeah, but I get this every time I go through TSA security, right? I’m being imaged. My face files are uploaded.

Your face, not your retina.

Yeah, but my face is probably—

Your iris.

Good enough.

Maybe. I mean, there’s a whole cottage industry of folks who look at the ability to deceive facial recognition with printouts or with 3D masks. So this is pushing it to the iris. But I think, for me, what the story underlines is that we’ve arrived early. That scene, that iconic scene in Minority Report set at the Gap, was set decades from now.

Right.

We caught up.

So let me get this right. I’m walking into—

This is the speed-running of every science-fiction story, all of them.

Right.

So I’m walking into the Gap, but before I can shop, I’ve got to stick my eyeball in the retinal reader, and then it’s going to serve me properly.

I think they have a 3-meter range on these things. I don’t know if these ones do, but the military has a 3-meter range on these.

It’ll get better, and you’ll be able to do it at a distance.

So, yeah, you just have to look in the direction.

Ah, you’ve got another glass of wine coming.

Good morning, dude.

All right. It’s going to increase the humor level. Fantastic. By the way, let me just take a second to thank the team who puts on Moonshots—Nick Singh, Dana Khan, and Gianluca—who do an amazing job every week supporting us. Can we give it up for that team?

Yeah, this is exciting news. On this stage about 2 years ago, I had Mike Andrec, the CEO of Aeon, which is one of your companies.

I think that was 1 year ago.

Was it 1 year? Man, oh, man.

1 year ago.

Okay, this feels—

Time compression.

Yeah.

Yeah.

But tell us about what Aion Systems is doing and what, in particular, you achieved here.

6. The Fruit Fly Brain Upload

Okay, so I think this ended up being the number-one technology story over the weekend, according to the various news feeds that I was seeing. So right here, actually—

Biased news feeds? No, just—

Yeah, of course. Right here, over the weekend, at the kickoff for this Abundance Summit, we announced—the Aeon Systems Public Benefit Corporation, that is—the first multi-behavior brain upload in the world, and this was of a fruit fly.

Aeon Systems, which I co-founded, has the goal of ultimately uploading human minds and nonhuman minds to cyberspace. We want to put a human in the cloud as soon as we possibly can.

This weekend, for the first time, we announced taking the brain of a fruit fly and putting together a few pieces that were really just sitting around. There was a bit of work from our senior scientist, Phil Xu, in 2024, looking at partial emulation of a fruit fly brain, and we put that together with a number of other models that were available—a mechatronic simulated model of a fruit fly and some other advances.

For the first time, we closed the sensorimotor arc: taking a fruit fly connectome, embedding that in a simulated world. You can see that in the video that’s playing here. Literally, I would say this is an early upload of a fruit fly. The fruit fly is able to walk around, scratch itself, and eat simulated banana.

At the same time, while on the left-hand side of the video you’re seeing the embodied experience showing multiple behaviors of the fruit fly, on the right-hand side, simultaneously, we’re modeling every single neuron in the fruit fly brain, and that’s driving the entire sensorimotor arc.

Wow.

50 million connections.

50 million.

50 million connections.

And it does not know it’s a fruit fly.

We don’t think the fruit fly knows that it’s a fruit fly.

Okay.

Not sure. This is an early experiment. I can’t emphasize how much of an early experiment this is, but hopefully history will regard this past weekend—it got a bunch of attention. Elon was excited by it, and others found it pretty exciting too—as the moment when the first model organism had an entire brain uploaded.

So what’s next? A mouse?

Yeah.

Yep, let’s give it up for this.

Well, clearly the next one has to be a lobster.

Lobster?

You said the lobster.

A lot of people asked that, right?

Isn’t that the plot of Accelerando?

Accelerando. Right. I can’t tell you how many people love to write and say, “You’re mispronouncing Accelerando. You have to pronounce it in the right Italian way,” which is Accelerando.

Okay, so for those who want it, Accelerando, yes, this is the plot point. We are speed-running every sci-fi trope everywhere, all at once, with Accelerando being one of those plot points.

Lobsters aren’t next. Aeon wants to go after mice, and it wants to go after humans, and we’re going to do this. Part of the reason why we want to do this is that right now the singularity, which I would argue we’re in the middle of, is filled with artificial minds.

The trillions of dollars in CapEx that we’re using to tile the Earth with compute is available only to artificial minds, to LLMs. It’s not available to any minds that, in any remote way—other than perhaps at the behavioral level—resemble human biological meat minds.

We want to level the playing field so that humanity can take advantage, on a level playing field, of the same compute advantage that right now is tipped in favor of these artificial minds, so we can put humanity into the cloud as well.

Amazing. 100 trillion synaptic connections for a human. How much for a mouse?

It’s orders of magnitude larger, and there’s some quibbling because it depends on how you measure the number of available weights or weight properties for synapses, and also how many cells—how many brain cells—end up being significant or not.

It’s orders of magnitude larger. This isn’t happening anytime soon. Just to anchor expectations appropriately, we don’t think we’re months away from a mouse or a human. But I think the right way to think about it is, at this point, it’s going to be years, not decades, before we get to the first mouse and the first human whole-brain emulations.

Amazing. Let’s move it to xAI. You know, it’s so funny. I’ve known Gwen Shotwell for 20 years now, and I’m so used to her reporting on Falcon and Dragon and rockets, and not xAI and gigawatt power centers.

We both, backstage, were like, “Why would Gwynne be talking about AI?”

Oh, yeah. SpaceX.

Right. They own it. I was like, “Ah.”

The Dyson swarm makes for strange bedfellows.

It does.

You know what blew my mind on this one? 1.2 gigawatts is about the energy used by the Dallas–Fort Worth metropolitan area.

So just to read this out, xAI has committed to develop 1.2 gigawatts of power as their supercomputer power source.

That’s one data center.

That’s unbelievable.

Per data center.

That will be with every additional data center. So every data center they build, they’re building at 1.2 gigawatts. So the question is, where are they going to get that from?

Well, this came up with Eric Schmidt, too. You remember we interviewed him last summer at your place, and he said, “We’re going to lose to China if we don’t find 100 gigawatts of power.” And then on the stage here yesterday, it was like, “Hey, what do you know? We're tracking to find the 100 billion. All we did was deregulate and put it in the hands of the companies. The companies are incredibly well-funded, and they’ll find the power because they care about their data centers actually operating.” And that’s how—

Well, what I find amazing as well is that this year, the U.S. is on target in 2026, I think, to add 86 gigawatts of new capacity to the grid, but 51% of that is solar.

To me, the power of the American entrepreneur is like nothing. It’s just mind-boggling to me that a guy like Sam Altman, who has nothing to do with the power industry, is going to say, “You know what? I’m going to find the gigawatts. I’m going to build nuclear reactors. I’m going into space.” It’s incredible.

Yeah.

The range of capability of an American entrepreneur when there’s a need is like no force in the world.

Let’s get to eVTOLs—flying cars. So, Florida advances a bill to formalize a regulatory flying-car framework. One of the things I’m proud of and excited about here in L.A. is that the L.A. Olympics are coming up.

Yeah.

And there’s Archer Aviation. The 2 major players in eVTOLs in the United States are Joby and Archer. There are other ones as well, but Archer plans to become operational by 2028 here and move people around different parts of Los Angeles because the traffic is going to suck. And we see a movement in Florida as well.

I’m just glad they didn’t say “Florida Man advances a bill,” because that would be a problem. But I think this is really important. The key word here for me is “framework,” because once you can start to set up the foundations for this, it means the whole model and the whole regulatory regime accelerates. And, God help us, we need this type of stuff yesterday.

Uh-huh.

Which I hope even Alex would agree. We don't have it yesterday.

I agree, but I also think we’re catching up with the future.

Yeah.

We’re finally getting flying cars, and I keep a mental bingo card of which sci-fi tropes we have not yet achieved in some fashion. We don’t have warp drive. Waiting for that one.

Yeah.

We don’t have yet—

Teleportation.

Teleportation, Star Trek replicators.

Time travel.

Time travel may or may not be physically possible, but it—

The replicator’s close. The holodeck is close. We’re very close to something.

We’re getting very close to a lot of sci-fi tropes.

All right.

Yeah.

7. The Abundance AMA

The fun part now is your questions. We’re going to do an AMA here with our Abundance community. As you know, let’s go to the mics. We’ll also entertain the questions from Zoom. All right, Christian. Let’s kick it off with you, buddy.

Thank you so much, Peter. Awesome to be here, guys. I watch you all the time, or I listen to you while I’m running. D.B. 2, awesome brother. Your insights. Imad, the guests are great. Peter, an awesome dream team.

He’s like our biggest fan.

Ismail, I’m glad, Saleem, that you got to check out that AWG is real—or at least in an android.

Finally.

I was suspicious for a long time.

Don’t believe it for a second. It’s just a meat body for rent.

He’s still an algorithm.

For now. My question is a little bit about the way that I get involved in this technology: it’s through a capitalist mindset. The word capital is really what constricts, and it’s been that way for maybe the last 200 or 300 years. And I keep getting this sensation that capital is getting less and less relevant—

Mm.

—and the idea of scarcity in economics from that Econ 101—the management of scarcity of services and needs—and scarcity is going more toward technology than capital. What kind of timelines are you guys looking at with this? I know it’s always a timeline question. Nobody has a crystal ball. But is there something that you guys are thinking about where we’re just going to get a little bit more and more squeezed out?

You know, I’ll give you one data point, because this came up on that last podcast we did, where Anthropic was saying they’re going to do about $26 billion in run rate, but they’re growing 10X year over year. And I did the math on the fly. I messed it up, of course, because I wasn’t Alex.

But if they grew 2 more years at 10X year over year, they go from $26 billion to $260 billion to $2.6 trillion—the most revenue in the history of the world. The PEG ratio implies that that company would be worth a quadrillion dollars.

Well, we heard Elon say we’re going to have $100 trillion companies, and I can imagine that within 5 years.

Yeah, so that would—

So 3 years from now, I mean—

That would mean 3 years.

A couple years. Yeah.

3 years.

I don’t think it’s going to be unreasonable. I mean, listen, it’s so funny the way all of a sudden a trillion here and a trillion there has become the accepted number.

Yeah.

I want to say something about this. A really, really key point that we’ve hit over the last couple of years is that innovation is not capital-constrained anymore.

Mm-hmm.

It used to be that you had an idea and your constraint was, could you go get funding for that idea? And so you had to go out to your investors, the VCs, the banks, and whatever, whatever. And it was only available in those places like Silicon Valley or Austin or whatever, where you had a preponderance of capital available.

We have today what we call PDI—permissionless disruptive innovation—where anybody can take on a very disruptive idea like Claude bot or take Vitalik Buterin, an 18-year-old kid out of Toronto, who ignores his professors and gets together with a few friends. Boom, you have a multihundred-billion-dollar ecosystem that nobody understands.

And so you have the opportunity today. It only comes down to mindset. And the reason, Peter, it’s so amazing that you run this event and put this community together is that the difference between the people in this world and the outside world is night and day, right? All of you have the problem that you go home to your family, your colleagues, whatever, and you cannot explain to them what happened, right? You’re like, “I can’t even process it.” You can’t make that gap.

So it only comes down to mindset now, which is the most amazing thing possible, because mindsets are fixable and shiftable.

So I had this little side conversation with Eric that you guys may have picked up, because I’ve had this conversation about whether we’re heading toward a post-capitalist society where money has very little value. And so what does have value in the future? And we’ve talked about this, Alex—it’s compute and energy, ultimately.

Did you ever read Zero Marginal Society?

No, I’m not sure I have.

By Jeremy Rifkin.

Huge. Yeah.

And it talks about where we’re going. Eventually, everything basically falls down to—

The marginal cost of production.

—the marginal cost, which is electricity, raw material—

It’s the inner loop.

—and data.

So if you want to build anything like an electric Ferrari, to use as an example, it’s the raw cost of it—the cost of extracting it—which drops in cost as you have—

Yeah, let me pick up on—

—robotic mining. Yeah.

Just for a second. Take 3D printing, right? It’s been around for a while. The big, profound breakthroughs in 3D printing are not that you can physically build something; it’s the fact that complexity becomes free.

Yes. And personalization becomes free.

In the past, complexity was expensive. The design, materials, and manufacturing capability of a complex object were more expensive than those of a simple object. But with 3D printing, complexity doesn’t matter. It doesn’t matter how complex the object is; it just builds it.

And as we get to molecular manufacturing, that goes to near zero again. So just those couple of breakthroughs across all of these domains, especially when you add AI as an accelerant to everything, mean that we have profound movement forward. Hence, we are in the middle of the singularity.

The one question I wish I had asked when we were with Elon, when he was talking about money having much less value, and I wanted to say, “So, just as you become a trillionaire, money has little value.”

You did ask that, didn’t you?

No, I didn’t. I didn’t.

It was off camera?

It was off camera.

Oh.

But I don't think it's a coincidence. I don't think that this is some cosmic irony that Elon is about to become a trillionaire at the same time that some folks—not including myself—are hand-wringing a bit that suddenly we're about to enter into some post-capitalist state where money becomes irrelevant. I think that this was always going to happen.

Agreed.

It was inevitable. And I just want to speak to what I understood the core of the question to be, which is that there's this cliché out there that capital fights labor, and capital usually wins, but this time around something different might happen. Historically, every time capitalism and labor get into a fight, capital usually wins. This time around, the risk is that maybe capital itself isn't immortal. Maybe capital is finally mortal for the first time in human history.

I'm not sure that that's the case. I think that would be, on the one hand, a nightmare scenario. On the other hand, Salim, you were talking about how we're entering some sort of post-scarce state, but arguably the trillions of dollars of CapEx that are going into tiling the Earth with compute, and soon solar synchronous orbit, and soon after that maybe the Dyson sphere—

Oh, the Moon.

Oh, damn it.

Drink, drink, drink, drink. Soon—even that—unless the physics of our universe turns out to be radically different, so radically different from what it looks like right now, I think there will probably always be certain scarce physical resources.

Of course.

It could look like control. It may or may not be energy; we'll see. It may or may not be the speed of light; we'll see. But to the extent that there are any scarce physical resources, and to the extent that there are ever multiple actors in the future, I think the laws of thermodynamics, and probably the laws of economics, will still apply.

We are still young as a species. Let's go to Ahmer on Zoom. Ahmer, good to see you. Pleasure. Welcome.

Good to see you as well. Thank you. Appreciate it. Happy to be here. Very quick question to the panelists. We are seeing Sam Altman raising $100 billion. Yann LeCun just raised $1 billion today to scale up world models. So are we still talking about scaling language models, or scaling physical simulation?

I'm curious what the panelists think about human intelligence and reasoning that goes much beyond just observation and language, and where you see the potential for true artificial intelligence evolving into superintelligent systems. Thank you.

Did you understand Ahmer's question?

It sounded a little bit—

I don't think exactly.

—like the stochastic parrot question, which is: Will we be able to generate new knowledge from these systems? I think the answer is—

You know, I think, having had some conversations with Ahmer, he's talking about symbolic AI and why we're not investing in symbolic AI.

Oh, you think this is the neurosymbolic question?

That's what I think it was.

Okay. Well, I'll offer my 2 cents. I'm sure you all have views as well. I think it's a false distinction. If this is the neurosymbolic question—why are we investing so much attention in LLMs and not in good old-fashioned AI, or symbolic discrete AI?—it's a totally false distinction.

We tokenize everything. I had an interesting discussion at Davos this year with Peter Danenberger from DeepMind, where we found ourselves in an interesting avenue, debating whether tokenization is a bit of a crime, a form of violence against knowledge, and whether discretization in general is doing harm to knowledge.

I think we need to bring you a couple of tequila shots here.

Yeah.

Let's go to Mark. Mark, go, please.

Earlier today, I challenged Dara from Uber to invest in the Abundance XPRIZE as an investor and a competitor to deliver housing, food, energy, and connectivity for $250 a month. We're investing $2 billion a day in compute and building data centers, a billion dollars a day in war, and I'm wondering what it's going to take to invest in people.

I want to put a larger challenge out today. I'm going to commit 1% of my wealth on an annual basis into a wealth fund, a small-scale pod of 44 people, 38 needs-based, and 7 or 8 contributors. It's going to distribute 5% per year: 4% goes as cash, and 1% goes to an expansion pool. You can read about it at markpatrickdonovan.com. And I'm challenging others to invest today, not tomorrow, and to mitigate this rough period. It doesn't have to be as rough if we put a fraction of what we're putting into compute into people.

Yeah.

We did that in Denver with the Denver Basic Income Project, where I leveraged $500,000 up to $10.8 million for people experiencing homelessness. And when you invest in people, it gives them hope. We need to do it today.

Yeah, Mark, I could not agree more. The challenge is that human nature is very egocentric and very self-centered. In other words, people are putting money where it's either meeting their immediate need or where it's going to give them more money in the long term.

You have to understand, if you look at philanthropy—which, by its definition, means “friend of man”—it's a very different pocket from the for-profit sector. I see this all the time because I'm raising money for my companies and raising money for my nonprofits, and the ratio, if you think about it, is between 100 to 1 and 1,000 to 1. For every dollar I donate, I'm willing to invest somewhere between $100 and $1,000, and that's what's out there right now. It's a challenge.

We are driven by fear, curiosity, and greed. I would posit that those are the 3 major human drivers. Love—you can add that as a potential fourth. Interestingly enough, you could measure the ratio of fear to curiosity. It's the ratio of the defense budget to the science budget, right? And greed is the ratio represented by the entire investment community.

Yeah. There's something very important in the work that you're doing with that XPRIZE. What we found with XPRIZE is that when you position a prize and launch it, it typically gets won within 6 to 7 years. It's a 10× drop from where we are today—about $2,500 a month—to where you're talking about, $250 a month, to pay for everything.

If we imagine that gets done in the next 6 to 7 years, it changes the equation globally, and it forces everybody to go, “Oh my God, that's possible.” When we get to that point, it'll completely change the game, especially as we get closer and can publicize the outcomes.

So this era of greed and ignoring the fundamental problems will literally disappear and evaporate in the next 2 to 3 years as we keep working that prize and getting the media word out there. This is incredibly powerful and important.

Peter and I, when we wrote this last book, wrote a section in there called “Technological Socialism.”

Mm.

Socialism—government socialism—fails because centralized allocation of assets is too inefficient and invariably leads to corruption. But if you think about Dara and the sharing of cars across a large group of people, it's actually a socialist application.

When an algorithm hyper-efficiently matches demand and supply, you get all the benefits of the sharing economy without the downsides, without the corruption, and without the inefficiencies. So we have all sorts of capabilities with algorithms and AI now to deliver much of what you're talking about in a hyper-efficient way. We just have to propagate those, and that's going to start to happen now.

Yeah. I wrote in my book, The Lost Economy, about this, and I've got a paper coming out soon where I look at the new monetary flows as agents basically crowding out the private sector.

My view is this: Everyone ultimately needs to have universal basic AI, or Claude, or whatever. That allows us to reach everyone. Everyone needs an AI that grows with them.

Money needs to come not from banks, but from being human. That's the only way the math works. It doesn't work from taxation; it doesn't work from anything else. You need that basic level of money coming into being, not from deposits at banks, but from being human. Then the AIs will buy from us, and that enables all of this with the AI that everyone has.

Mm-hmm. Professor Brown.

We had half a day of really interesting talks whose subtext was massive job loss. Then we had another half a day of talks about massive labor scarcity, which is why we need all these robots. So, aside from temporary displacements, which we know are going to happen, which is it?

It's clearly a massive trough, massive social unrest, and then a rebound in 2028. It was interesting to hear Eric backstage come up with basically the same timeline. But it's almost like the Industrial Revolution all over again, except instead of over 20, 30, or 40 years, it's over 2, 3, or 4 years.

And so a huge amount of retooling needs to happen. The way we do taxation and government needs to get restructured. All of that is gonna—AI's just gonna happen way too quickly for all those things to react. But then a massive amount of unrest, and then 2028, hopefully it will deploy again.

I have the counterpoint. I don't think we're gonna see massive job loss because I think what's gonna happen—I’m writing a paper right now called “The Organizational Singularity,” right? Because as agents take over all execution, even strategy inside companies essentially dissolves to the work of AI. So what do you do?

The calculations we've done so far indicate that you'll take a typical company, automate everything with AI, and you'll end up with about 25% of the same number of people doing oversight, managing dashboards, doing exception handling, and owning the purpose of the organization, okay? But you end up creating 5 times more companies because you can, and therefore, the employment stays exactly the way it has.

And this is what we've seen consistently throughout history, where we have a disruption, but all sorts of other sectors take up the slack, and we don't end up with radical unemployment. So I tend to be much more optimistic. Take my veil off—

I just know—

Call it the wine, call it whatever. But I tend to be much more optimistic than—

All right. I'm gonna move this forward because it's past my bedtime. We go to Brad, and then we go to Pete, and we're gonna wrap it there. Brad, please go ahead.

Wait, I want to—

Yes.

After we finish, I do want some commentary from the group—

We'll—

You folks—

I'll take care of that. Brad.

Salim, I'm gonna give you an assist here. And maybe this is a topic for your talk late tomorrow night, but maybe Moltbook is an example of, in this age of artificial intelligence, the rising value of ingenuity and creativity.

Maybe what Meta acquired was not strategic, and we're all overthinking it, and they just liked the team. They thought that they were creative, that they had some sort of magic, and they wanted to capture that magic inside their company, and that's why it was acquired.

So I just wanna capture your thoughts, this great minds up on the stage there, on the rise of ingenuity and creativity and the value of that.

Yeah.

I think we're way overthinking this. I've got 1,100 people, and I know them firsthand, and many of them I genuinely love. Lots of them have been in the same roles for 10 or 15 years. They're great at it, they've perfected it, and then AI just comes along one day, and it can do it.

And there's huge pressure on the management team for higher margins, higher profits. So what's gonna happen is obvious. The valuations of the companies are gonna go through the roof, so their shareholders will make a lot more money, but their W-2 paycheck is dead. It's going away, and it's gonna create—

It's cooked.

—a huge amount of disruption.

It's cooked.

Some subset of people are shareholders. All my people are shareholders, so they'll be okay. Lots of other people are not shareholders. All Dara's drivers are not shareholders, I think, as far as I know. So they're in deep trouble.

The idea that somehow they're gonna become creators overnight is ludicrous. The people who are creative, like the Moltbook people, they're gonna do incredibly well. Our kids, and most kids who are not saddled by a career, are gonna do incredibly well.

But in transition, it's inevitable. It's happening imminently.

All right, Pete. We'll just—

Wait, wait. I want one statement.

All right. Okay.

What we found with the Exponential Organizations model is that survival and success depend on adaptability, not scalability and efficiency. And so you just keep that vector going. The people who are the most adaptable today, they're gonna survive the most.

Amen. Throw your kids into the woods and see if they survive.

No, I didn't say that.

Alex, I said this to you earlier today. I love your analogy of tiling the planet with compute because, as a data center design builder, my answer to the power problem is that finding 1,200 megawatts of contiguous property is getting harder and harder.

So my answer is that's only 120 10-megawatt data centers, and you put them in an area, and we tile the areas to be able to do that. And Ahmad, I think it matches perfectly with your idea of national champions because what you're trying to do for the protocol stack of decentralization and sovereignty, I wanna do at the physical layer.

I wanna build 20,000 data centers across the country at 10 megawatts so that I'm in less than 1 millisecond from any place in the country, if you will—the high school football cities of the world.

That's huge.

And to me, that solves it on both sides: at the protocol layer and from the data center distribution standpoint. And I think that's how we can actually deliver the power, because we don't have a power production problem in this country. We have a power transmission and storage problem in this country.

And I think every governor in the country should hear exactly what you just said and jump on it instantly. And Alex is incredibly frustrated with the meetings we've had with government.

Shh.

Look, if you're right, and I hope you are, and I think you probably are, then we need lots and lots of regional data centers that have to be in every single state, and that would be the best thing that could ever happen for this job dislocation.

So if that theory is right, we need to get on it right away and create those projects, like now.

I'm ready.

All right. Let's give it up for Alex Wissner-Gross, Dave Blundin, Salim Ismail, and Imad Mushtaq. Whoo.